BERT for TensorFlow2

The performance measurements in this document were conducted at the time of publication and may not reflect the performance achieved from NVIDIA's latest software release. For the most up-to-date performance measurements, go to NVIDIA Data Center Deep Learning Product Performance.

Benchmarking

The following section shows how to run benchmarks measuring the model performance in training and inference modes.

Both of these benchmarking scripts enable you to run a number of epochs, extract performance numbers, and run the BERT model for fine tuning.

Training performance benchmark

Training benchmarking can be performed by running the script:

scripts/finetune_train_benchmark.sh <bert_model> <num_gpu> <batch_size> <precision> <use_xla>

This script runs 800 steps by default on the SQuAD v1.1 dataset and extracts performance numbers for the given configuration. These numbers are saved at /results/squad_train_benchmark_<bert_model>_gpu<num_gpu>_bs<batch_size>.log.

Inference performance benchmark

Inference benchmarking can be performed by running the script:

scripts/finetune_inference_benchmark.sh <bert_model> <batch_size> <precision> <use_xla>

This script runs 1000 eval iterations by default on the SQuAD v1.1 dataset and extracts performance and latency numbers for the given configuration. These numbers are saved at /results/squad_inference_benchmark_<bert_model>_<precision>_bs<batch_size>.log.

Results

The following sections provide details on how we achieved our performance and accuracy in training and inference for fine tuning Question Answering. All results are on BERT-Large model unless otherwise mentioned. All fine tuning results are on SQuAD v1.1 using a sequence length of 384 unless otherwise mentioned.

Training accuracy results

Pre-training accuracy

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 and NVIDIA DGX A100.

DGX System	Nodes x GPUs	Precision	Batch Size/GPU: Phase1, Phase2	Accumulation Steps: Phase1, Phase2	Time to Train (Hrs)	Final Loss
DGX2H	32 x 16	FP16	56, 10	2, 6	2.67	1.69
DGX2H	32 x 16	FP32	32, 4	4, 16	8.02	1.71
DGXA100	32 x 8	FP16	312, 40	1, 3	2.02	1.68
DGXA100	32 x 8	TF32	176, 22	2, 6	3.57	1.67

Fine-tuning accuracy for SQuAD v1.1: NVIDIA DGX A100 (8x A100 40GB)

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.12-py3 NGC container on NVIDIA DGX A100 with 8x A100 80GB GPUs.

GPUs	**Batch size / GPU: TF32, FP16 **	Accuracy - TF32	Accuracy - mixed precision	Time to Train - TF32 (Hrs)	Time to Train - mixed precision (Hrs)
8	38, 76	90.88	91.12	0.16	0.11

Pre-training SQuAD v1.1 stability test: NVIDIA DGX A100 (256x A100 80GB)

The following tables compare Final Loss scores across 3 different training runs with different seeds, for both FP16 and TF32. The runs showcase consistent convergence on all 3 seeds with very little deviation.

FP16, 256x GPUs	seed 1	seed 2	seed 3	mean	std
Final Loss	1.657	1.661	1.683	1.667	0.014

TF32, 256x GPUs	seed 1	seed 2	seed 3	mean	std
Final Loss	1.67	1.654	1.636	1.653	0.017

Fine-tuning SQuAD v1.1 stability test: NVIDIA DGX A100 (8x A100 80GB)

The following tables compare F1 scores across 5 different training runs with different seeds, for both FP16 and TF32 respectively using the (NVIDIA Pretrained Checkpoint)[https://ngc.nvidia.com/catalog/models]. The runs showcase consistent convergence on all 5 seeds with very little deviation.

FP16, 8x GPUs	seed 1	seed 2	seed 3	seed 4	seed 5	mean	std
F1	91.12	90.80	90.94	90.90	90.94	90.94	0.11

TF32, 8x GPUs	seed 1	seed 2	seed 3	seed 4	seed 5	mean	std
F1	90.79	90.88	90.80	90.88	90.83	90.84	0.04

Training performance results

Pre-training training performance: Single-node on NVIDIA DGX-2 V100 (16x V100 32GB)

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs	Sequence Length	Batch size / GPU: mixed precision, FP32	Gradient Accumulation: mixed precision, FP32	Global Batch Size: mixed precision, FP32	Throughput - mixed precision	Throughput - FP32	Throughput speedup (FP32 - mixed precision)	Weak scaling - mixed precision	Weak scaling - FP32
1	128	60 , 32	1024 , 2048	61440 , 65536	206.5	49.97	4.13	1.00	1.00
4	128	60 , 32	256 , 512	61440 , 65536	789.75	194.02	4.07	3.82	3.88
8	128	60 , 32	128 , 256	61440 , 65536	1561.77	367.9	4.25	7.56	7.36
16	128	60 , 32	64 , 128	61440 , 65536	3077.99	762.22	4.04	14.9	15.25
1	512	10 , 6	3072 , 5120	30720 , 30720	40.95	11.06	3.70	1.00	1.00
4	512	10 , 6	768 , 1280	30720 , 30720	158.5	43.05	3.68	3.87	3.89
8	512	10 , 6	384 , 640	30720 , 30720	312.03	85.51	3.65	7.62	7.73
16	512	10 , 4	192 , 512	30720 , 32768	614.94	161.38	3.81	15.02	14.59

Note: The respective values for FP32 runs that use a batch size of 60 and 10 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.

Pre-training training performance: Multi-node on NVIDIA DGX-2H V100 (16x V100 32GB)

Our results were obtained by running the run.sub training script in the TensorFlow 21.02-py3 NGC container using multiple NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

Num Nodes	Sequence Length	Batch size / GPU: mixed precision, FP32	Gradient Accumulation: mixed precision, FP32	Global Batch Size: mixed precision, FP32	Throughput - mixed precision	Throughput - FP32	Throughput speedup (FP32 - mixed precision)	Weak scaling - mixed precision	Weak scaling - FP32
1	128	60 , 32	64 , 128	61440 , 65536	3528.51	841.72	4.19	1.00	1.00
4	128	60 , 32	16 , 32	61440 , 65536	13370.21	3060.49	4.37	3.79	3.64
16	128	60 , 32	4 , 8	61440 , 65536	42697.42	10383.57	4.11	12.1	12.34
32	128	60 , 32	2 , 4	61440 , 65536	84223.16	20094.14	4.19	23.87	23.87
1	512	10 , 4	192 , 256	30720 , 32768	678.35	180	3.77	1.00	1.00
4	512	10 , 4	96 , 64	30720 , 32768	2678.29	646.76	4.14	3.95	3.59
16	512	10 , 4	24 , 32	30720 , 32768	7834.72	2204.72	3.55	11.55	12.25
32	512	10 , 4	6 , 16	30720 , 32768	18786.93	4196.15	4.48	27.70	23.31

Note: The respective values for FP32 runs that use a batch size of 60 and 10 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.

Pre-training training performance: Single-node on NVIDIA DGX A100 (8x A100 80GB)

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX A100 with 8x A100 80GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs	Sequence Length	Batch size / GPU: mixed precision, TF32	Gradient Accumulation: mixed precision, TF32	Global Batch Size: mixed precision, FP32	Throughput - mixed precision	Throughput - TF32	Throughput speedup (TF32 - mixed precision)	Weak scaling - mixed precision	Weak scaling -TF32
1	128	312 , 176	256 , 512	79872 , 90112	485.59	282.98	1.72	1.00	1.00
8	128	312 , 176	32 , 64	79872 , 90112	3799.24	1944.77	1.95	7.82	6.87
1	512	40 , 22	768 , 1536	30720 , 33792	96.52	54.92	1.76	1.00	1.00
8	512	40 , 22	96 , 192	30720 , 33792	649.69	427.39	1.52	6.73	7.78

Note: The respective values for TF32 runs that use a batch size of 312 and 40 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.

Pre-training training performance: Multi-node on NVIDIA DGX A100 (8x A100 80GB)

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Performance (in sentences per second) is the steady state throughput.

Num Nodes	Sequence Length	Batch size / GPU: mixed precision, TF32	Gradient Accumulation: mixed precision, TF32	Global Batch Size: mixed precision, FP32	Throughput - mixed precision	Throughput - TF32	Throughput speedup (TF32 - mixed precision)	Weak scaling - mixed precision	Weak scaling -TF32
1	128	312 , 176	32 , 64	79872 , 90112	3803.82	2062.98	1.84	1.00	1.00
2	128	312 , 176	16 , 32	79872 , 90112	7551.37	4084.76	1.85	1.99	1.98
8	128	312 , 176	4 , 8	79872 , 90112	29711.11	16134.02	1.84	7.81	7.82
32	128	312 , 176	1 , 2	79872 , 90112	110280.73	59569.77	1.85	28.99	28.88
1	512	40 , 22	96 , 192	30720 , 33792	749.73	431.89	1.74	1.00	1.00
2	512	40 , 22	48 , 96	30720 , 33792	1491.87	739.14	2.02	1.99	1.71
8	512	40 , 22	12 , 24	30720 , 33792	5870.83	2926.58	2.01	7.83	6.78
32	512	40 , 22	3 , 6	30720 , 33792	22506.23	11240.5	2.00	30.02	26.03

Note: The respective values for TF32 runs that use a batch size of 312 and 40 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.

Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (8x V100 16GB)

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Performance (in sentences per second) is the mean throughput from 2 epochs.

GPUs	Batch size / GPU: mixed precision, FP32	Throughput - mixed precision	Throughput - FP32	Throughput speedup (FP32 to mixed precision)	Weak scaling - FP32	Weak scaling - mixed precision
1	6,3	39.10	9.85	3.97	1.00	1.00
4	6,3	128.48	36.52	3.52	3.29	3.71
8	6,3	255.36	73.03	3.5	6.53	7.41

Note: The respective values for FP32 runs that use a batch size of 6 are not available due to out of memory errors that arise. Batch size of 6 is only available on using FP16.

To achieve these same results, follow the Quick Start Guide outlined above.

Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (8x V100 32GB)

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 8x V100 32GB GPUs. Performance (in sentences per second) is the mean throughput from 2 epochs.

GPUs	Batch size / GPU: mixed precision, FP32	Throughput - mixed precision	Throughput - FP32	Throughput speedup (FP32 to mixed precision)	Weak scaling - FP32	Weak scaling - mixed precision
1	12,8	47.06	11.11	4.24	1.00	1.00
4	12,8	165.26	42.84	3.86	3.51	3.86
8	12,8	330.29	85.91	3.84	7.02	7.73

Note: The respective values for FP32 runs that use a batch size of 12 are not available due to out of memory errors that arise. Batch size of 12 is only available on using FP16.

To achieve these same results, follow the Quick Start Guide outlined above.

Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX A100 (8x A100 80GB)

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the mean throughput from 2 epochs.

GPUs	Batch size / GPU: mixed precision, TF32	Throughput - mixed precision	Throughput - FP32	Throughput speedup (FP32 to mixed precision)	Weak scaling - FP32	Weak scaling - mixed precision
1	76,38	134.22	43.9	3.057	1.00	1.00
8	76,38	1048.23	341.31	3.071	7.81	7.77

Note: The respective values for TF32 runs that use a batch size of 76 are not available due to out of memory errors that arise. Batch size of 12 is only available on using FP16.

To achieve these same results, follow the Quick Start Guide outlined above.

Inference performance results

Fine-tuning inference performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (1x V100 16GB)

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 1x V100 16GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

BERT-LARGE FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	105.04	1.277237354	9.52	9.67	9.77	10.16
128	2	184.9	1.671487977	10.82	11.15	11.27	11.8
128	4	301.9	2.448102498	13.25	13.38	13.45	13.96
128	8	421.98	3.149809659	18.96	19.12	19.2	19.82
384	1	74.99	2.15055922	13.34	13.5	13.58	14.53
384	2	109.84	2.709422792	18.21	18.4	18.6	19.39
384	4	142.58	3.313502208	28.05	28.28	28.48	28.85
384	8	168.34	3.823302294	47.52	47.74	47.86	48.52

BERT-Large FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	82.24	12.16	12.28	12.33	12.92
128	2	110.62	18.08	18.22	18.28	18.88
128	4	123.32	32.44	32.72	32.82	32.98
128	8	133.97	59.71	60.29	60.49	60.69
384	1	34.87	28.67	28.92	29.02	29.33
384	2	40.54	49.34	49.74	49.86	50.05
384	4	43.03	92.97	93.59	93.75	94.57
384	8	44.03	181.71	182.34	182.48	183.03

BERT-Base FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	236.26	1.179589595	4.23	4.37	4.49	4.59
128	2	425.1	1.441554478	4.7	4.84	4.97	5.26
128	4	710.48	1.911691107	5.63	5.78	5.93	6.4
128	8	1081.17	2.523032764	7.4	7.5	7.54	7.73
384	1	190.53	1.757170525	5.25	5.35	5.42	5.8
384	2	289.67	2.248292456	6.9	7.08	7.24	7.57
384	4	404.03	2.946328302	9.9	10	10.03	10.13
384	8	504.24	3.450153951	15.87	15.96	16.01	16.3

BERT-Base FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	200.29	4.99	5.08	5.16	5.53
128	2	294.89	6.78	6.89	6.93	7.37
128	4	371.65	10.76	10.89	10.96	11.92
128	8	428.52	18.67	18.89	18.98	19.17
384	1	108.43	9.22	9.26	9.31	10.24
384	2	128.84	15.52	15.6	15.71	16.49
384	4	137.13	29.17	29.4	29.48	29.64
384	8	146.15	54.74	55.19	55.3	55.54

To achieve these same results, follow the Quick Start Guide outlined above.

Fine-tuning inference performance for SQuAD v1.1 on NVIIDA DGX-1 V100 (1x V100 32GB)

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 1x V100 32GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

BERTLarge FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	101.58	1.242112986	9.84	9.99	10.06	10.39
128	2	181.89	1.651593571	11	11.14	11.2	11.87
128	4	295.86	2.348840902	13.52	13.67	13.75	14.5
128	8	411.29	3.010246652	19.45	19.62	19.69	20.4
384	1	72.95	2.083690374	13.71	13.93	14.08	14.81
384	2	107.02	2.583775954	18.69	18.8	18.88	19.57
384	4	139.8	3.14652262	28.61	28.75	28.88	29.6
384	8	163.68	3.595782074	48.88	49.09	49.18	49.77

BERT-Large FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	81.78	12.23	12.37	12.43	13.2
128	2	110.13	18.16	18.29	18.37	19.27
128	4	125.96	31.76	32.09	32.21	32.42
128	8	136.63	58.55	58.93	59.05	59.4
384	1	35.01	28.56	28.81	28.94	29.16
384	2	41.42	48.29	48.57	48.67	49.02
384	4	44.43	90.03	90.43	90.59	90.89
384	8	45.52	175.76	176.66	176.89	177.33

BERT-Base FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	234.85	1.217533309	4.26	4.33	4.37	4.62
128	2	415.86	1.435782351	4.81	4.92	5.06	5.55
128	4	680.09	1.84912586	5.88	6.1	6.2	6.53
128	8	1030.03	2.264548752	7.77	7.87	7.95	8.53
384	1	183.18	1.700993593	5.46	5.56	5.61	5.93
384	2	275.77	2.175528558	7.25	7.38	7.44	7.89
384	4	385.61	2.778570399	10.37	10.56	10.63	11.1
384	8	488.45	3.292329469	16.38	16.48	16.52	16.64

BERT-Base FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	192.89	5.18	5.3	5.36	5.65
128	2	289.64	6.91	7	7.22	7.83
128	4	367.79	10.88	10.98	11.02	11.59
128	8	454.85	17.59	17.76	17.81	17.92
384	1	107.69	9.29	9.37	9.42	9.88
384	2	126.76	15.78	15.89	15.97	16.72
384	4	138.78	28.82	28.98	29.06	29.88
384	8	148.36	53.92	54.16	54.26	54.58

To achieve these same results, follow the Quick Start Guide outlined above.

Fine-tuning inference performance for SQuAD v1.1 on NVIDIA DGX A100 (1x A100 80GB)

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 1x V100 32GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

BERT-Large FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	145.21	0.9435347628	6.89	7.14	7.4	8.35
128	2	272.81	1.093953003	7.33	7.61	7.77	8.35
128	4	468.98	1.273087573	8.53	8.71	8.83	9.85
128	8	705.67	1.191627687	11.34	11.64	11.9	13.1
384	1	118.34	1.042459479	8.45	8.82	8.99	9.52
384	2	197.8	1.231478023	10.11	10.48	10.62	11.4
384	4	275.19	1.268332027	14.54	14.73	14.8	16.8
384	8	342.22	1.416004634	23.38	23.64	23.75	24.1

BERT-Large TF32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	153.9	6.5	6.76	6.86	7.4
128	2	249.38	8.02	8.22	8.34	9.45
128	4	368.38	10.86	11.11	11.24	12.76
128	8	592.19	13.51	13.64	13.77	15.85
384	1	113.52	8.81	9.02	9.16	10.19
384	2	160.62	12.45	12.61	12.68	14.47
384	4	216.97	18.44	18.6	18.7	18.84
384	8	241.68	33.1	33.29	33.36	33.5

BERT-Base FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	295.01	1.014023992	3.39	3.59	3.65	3.73
128	2	594.81	1.048455898	3.36	3.59	3.68	4.19
128	4	1043.12	1.005145599	3.83	3.97	4.2	4.44
128	8	1786.25	1.198278638	4.48	4.73	4.8	5.19
384	1	278.85	1.103395062	3.59	3.67	3.99	4.15
384	2	464.77	1.252006896	4.3	4.59	4.87	5.29
384	4	675.82	1.264822578	5.92	6.15	6.27	6.94
384	8	846.81	1.31109494	9.45	9.65	9.74	11.03

BERT-Base TF32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	290.93	3.44	3.61	3.73	4.69
128	2	567.32	3.53	3.64	3.96	5.01
128	4	1037.78	3.85	3.95	4.06	4.58
128	8	1490.68	5.37	5.61	5.66	6.19
384	1	252.72	3.96	3.96	4.52	4.66
384	2	371.22	5.39	5.64	5.71	6.38
384	4	534.32	7.49	7.69	7.76	8.56
384	8	645.88	12.39	12.61	12.67	12.77

To achieve these same results, follow the Quick Start Guide outlined above.

Fine-tuning inference performance for SQuAD v1.1 on NVIDIA Tesla T4 (1x T4 16GB)

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 21.02-py3 NGC container on NVIDIA Tesla T4 with 1x T4 16GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

BERT-Large FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	57.6	1.364605544	17.36	18.16	19.02	21.67
128	2	102.76	2.17988969	19.46	20.68	21.27	22.2
128	4	151.11	3.146813828	26.47	26.9	27.06	27.45
128	8	186.99	3.733080455	42.78	43.87	44.18	44.78
384	1	38.88	2.590273151	25.72	26.06	26.16	26.38
384	2	50.53	3.202154626	39.58	39.93	40.35	40.95
384	4	57.69	3.721935484	69.34	70.5	70.77	71.09
384	8	62.99	3.927057357	127	129.18	130.07	131.86

BERT-Large FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	42.21	23.69	24.8	25.02	25.48
128	2	47.14	42.42	43.48	43.63	44.32
128	4	48.02	83.29	84.37	84.68	85.14
128	8	50.09	159.72	161.66	161.97	162.52
384	1	15.01	66.63	67.76	68.08	68.66
384	2	15.78	126.78	128.21	128.58	129.08
384	4	15.5	258.1	261.01	261.66	262.55
384	8	16.04	498.61	504.29	504.74	505.55

BERT-Base FP16

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Throughput speedup (FP32 to mixed precision)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	116.56	1.039878669	8.58	9.53	10.84	11.74
128	2	238.62	1.675937632	8.38	9.09	9.27	12.33
128	4	402.93	2.440964439	9.93	10.07	10.13	12.17
128	8	532.56	3.052619512	15.02	15.43	15.6	16.52
384	1	102.12	2.035073735	9.79	11.06	11.18	12.07
384	2	149.3	2.910898811	13.4	13.54	13.62	14.36
384	4	177.78	3.563439567	22.5	23.11	23.27	23.59
384	8	192.61	3.752386519	41.53	42.67	42.81	43.31

BERT-Base FP32

Sequence Length	Batch Size	Throughput-Average(sent/sec)	Latency-Average(ms)	Latency-90%(ms)	Latency-95%(ms)	Latency-99%(ms)
128	1	112.09	8.92	9.12	9.49	10.93
128	2	142.38	14.05	14.34	14.48	15.03
128	4	165.07	24.23	24.86	24.92	25.05
128	8	174.46	45.86	46.71	46.8	47.2
384	1	50.18	19.93	20.53	21.04	21.73
384	2	51.29	38.99	39.68	39.93	40.2
384	4	49.89	80.18	81.54	82	82.65
384	8	51.33	155.85	158.11	158.5	159.17

To achieve these same results, follow the Quick Start Guide outlined above.